IP header only is good enough for me. I tried increasing the flows to 1.1.1.[10-20]. Same result. The same flows and the same testbed distributes evenly on 82580 controllers with 5.3.2 and I211 on 4.1.2
On Wed, Dec 16, 2015 at 7:27 PM, Alexander Duyck <alexander.du...@gmail.com> wrote: > On Wed, Dec 16, 2015 at 8:11 AM, Tal Abudi <talab...@gmail.com> wrote: > > Hi > > I have an issue with igb 5.3.2 multi queue on a 2.6.18 machine with I211 > > controllers. > > When RSS is enabled (2 RX queues) all packets arrive to rx_queue_0 only. > > I'm running 2 UDP traffic streams (1.1.1.[10,11]:1024 -> 2.2.2.10:1025). > > > > With igb 4.3.2 traffic distributed evenly on all queues. > > The same configuration works great on a 82580 controller. > > > > ethtool -S enp5s0 | grep rx_q.*pa > > rx_queue_0_packets: 1561216274 > > rx_queue_1_packets: 0 > > > > Is this a known limitations ? > > Any tips ? > > Thanks ! > > The problem is that only 2 flows is a fairly small sample set when you > only have 2 queues. The odds of all flows ending up on the same CPU > are 2 / (2 ^ n). So in the case of 2 flows you have a 50% chance that > the hashes end up working out such that they all end up on the same > CPU. If you increase the number of flows the likelihood of the flows > being distributed increases. > > Also as you may be aware RSS only uses the IP headers unless you > enable them via the command "ethtool -N <ethX> rx-flow-hash udp4 > sdfn", however by doing that you can cause out-of-order packets for > fragmented flows. > > - Alex > -- Best regards, Tal Abudi
------------------------------------------------------------------------------
_______________________________________________ E1000-devel mailing list E1000-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/e1000-devel To learn more about Intel® Ethernet, visit http://communities.intel.com/community/wired
